How to make a SFINAE-based Y combinator in C++? - c++

I was thinking about the implicit templates of C++14, and I'm trying to declare a function to match an specific argument type (SFINAE and traits still give me headaches). I'm not sure how to explain what I want, but I'm trying to make a Y combinator (just to see if it's possible, not intended for production).
I'm trying to declare a function:
template<typename T>
my_traits<T>::return_type Y(T t) {
// ...
};
Such that T is a function (or a functor) that matches
std::function<R(F, Args...)>
// where F (and above return_type) will be
std::function<R(Args...)>
Which would take any number of arguments, but the first should be a function with the same return type and the same arguments (except this function itself). The first parameter to the operator () of the functor is a template.
The usage I want to achieve:
auto fib = [](auto myself, int x) {
if(x < 2)
return 1;
return myself(x - 1) + myself(x - 2);
};
// The returned type of fib should be assignable to std::function<int(int)>
I wasn't able to take the return type of the T type (because of the overloaded operator ()). What I'm trying to make is possible? How could I make it?
Edit:
Seeing it from a different angle, I'm trying to make this work:
struct my_functor {
template<typename T>
char operator () (T t, int x, float y) { /* ... */ };
};
template<typename T>
struct my_traits {
typedef /* ... */ result_type;
/* ... */
};
// I want this to be std::function<char(int, float)>, based on my_functor
using my_result =
my_traits<my_functor>::result_type;

It is not possible in C++14 return type deduction to deduce int(int) out of int(T, int) as OP desires.
However, we can mask the first parameter of the result using the following approach. The struct YCombinator is instantiated with a non-recursive function object member, whose first argument is a version of itself without the first argument. YCombinator provides a call operator that receives the arguments of the non-recursive function and then returns its function object member after substituting itself for the first argument. This technique allows the programmer to avoid the messiness of myself(myself, ...) calls within the definition of the recursive function.
template<typename Functor>
struct YCombinator
{
Functor functor;
template<typename... Args>
decltype(auto) operator()(Args&&... args)
{
return functor(*this, std::forward<Args>(args)...);
}
};
A make_YCombinator utility template allows for a streamlined usage pattern. This compiles run runs in GCC 4.9.0.
template<typename Functor>
decltype(auto) make_YCombinator(Functor f) { return YCombinator<Functor> { f }; }
int main()
{
auto fib = make_YCombinator([](auto self, int n) -> int { return n < 2 ? 1 : self(n - 1) + self(n - 2); });
for (int i = 0; i < 10 ; ++i)
cout << "fib(" << i << ") = " << fib(i) << endl;
return 0;
}
Since the non-recursive function is not defined at time that the recursive function is defined, in general the recursive function must have an explicit return type.
Edit:
However, it may be possible for the compiler to deduce the return type in certain cases if the programmer takes care to indicate the return type of the recursive function before use of the non-recursive function. While the above construction requires an explicit return type, in the following GCC 4.9.0 has no problem deducing the return type:
auto fib = make_YCombinator([](auto self, int n) { if (n < 2) return 1; return self(n - 1) + self(n - 2); });
To pin this down just a bit further, here is a quote from the draft C++14 standard on return type deduction [7.1.6.4.11]:
If the type of an entity with an undeduced placeholder type is needed
to determine the type of an expression, the program is ill-formed.
Once a return statement has been seen in a function, however, the
return type deduced from that statement can be used in the rest of the
function, including in other return statements. [ Example:
auto n = n; // error, n’s type is unknown
auto f();
void g() { &f; } // error, f’s return type is unknown
auto sum(int i) {
if (i == 1)
return i; // sum’s return type is int
else
return sum(i-1)+i; // OK, sum’s return type has been deduced
}
—end example ]

It's a really hacky approach, and has severe limitations, but here it goes:
First, we need a class that pretends to support every possible operation (as far as possible), such as the fake_anything class. Note that this isn't perfect since at a minimum . and :: won't work. To fake a functor, we give it a function call operator:
template<class... Ts> fake_anything operator()(Ts&&...) const;
Knowing that the lambda has only one operator(), and that operator() has only one template parameter allows us to extract its signature with decltype(&T::operator()<fake_anything>).
For this to work, the lambda's return type must be explicitly specified; it can't use deduction, since otherwise the deduced return types will probably conflict.
Finally we can obtain the other arguments to the lambda and the return type using the standard partial specialization approach:
template<class T>
struct extract_signature;
template<class T, class R, class FA, class...Args>
struct extract_signature<R (T::*)(FA, Args...)> {
static_assert(std::is_same<fake_anything, std::decay_t<FA>>::value, "Unexpected signature");
using type = std::function<R(Args...)>;
};
template<class T, class R, class FA, class...Args>
struct extract_signature<R (T::*)(FA, Args...) const> {
static_assert(std::is_same<fake_anything, std::decay_t<FA>>::value, "Unexpected signature");
using type = std::function<R(Args...)>;
};
// other cv- and ref-qualifier versions omitted - not relevant to lambdas
// we can also static_assert that none of Args is fake_anything, or reference to it, etc.
And add an alias template to hide all the ugliness of the hack:
template<class T>
using signature_t = typename extract_signature<decltype(&T::template operator()<fake_anything>)>::type;
And finally we can check that
static_assert(std::is_same<signature_t<decltype(fib)>,
std::function<int(int)>>::value, "Oops");
Demo.
The limitations:
The return type of operator() must be explicitly specified. You cannot use automatic return type deduction, unless all of the return statements return the same type regardless of the return type of the functor.
The faking is very imperfect.
This works for operator() of a particular form only: template<class T> R operator()(T, argument-types...) with or without const, where the first parameter is T or a reference to possibly cv-qualified T.

Related

return type deduction of recursive function

Recently, I read Barry's answer to this question Recursive lambda functions in C++11:
template <class F>
struct y_combinator {
F f; // the lambda will be stored here
// a forwarding operator():
template <class... Args>
decltype(auto) operator()(Args&&... args) const {
// we pass ourselves to f, then the arguments.
// [edit: Barry] pass in std::ref(*this) instead of *this
return f(std::ref(*this), std::forward<Args>(args)...);
}
};
// deduction guide
template <class F> y_combinator(F) -> y_combinator<F>;
Basically, y_combinator allows one to write a recursive lambda expression more easily (e.g. without having to delcare a std::function). When I played with y_combinator, I found something strange:
int main() {
// Case #1 compiles fine
y_combinator{[](auto g, int a, int b) {
if (a >= b) return 0;
return 1 + g(a + 1, b);
}}(1, 2);
// Case #2 deos not compile
y_combinator{[](auto g, int a) {
if (a >= 0) return 0;
return 1 + g(a + 1);
}}(1);
// Case #3 compiles just fine
y_combinator{[](auto g, int a)->int {
if (a >= 0) return 0;
return 1 + g(a + 1);
}}(1);
}
Case #1 and Case #3 compile fine while Case #2 does not compile. I got the same result with Clang 10.0 and GCC 9.3.
For Case #2, Clang says
prog.cc:25:18: error: no matching function for call to object of type 'std::__1::reference_wrapper<const y_combinator<(lambda at prog.cc:23:18)> >'
return 1 + g(a + 1);
^
How is the different results between Case #1 and Case #2?
Why does the trailing return type make a difference between Case #2 and Case #3?
You can check it on Wandbox.
The difference is that in #1 the initial and recursive calls to y_combinator have different argument types, whereas in #2 they have the same argument types (including value category).
In #1, the initial arguments (1, 2) are both int prvalue, whereas the recursive arguments g(a + 1, b) are respectively int prvalue and int lvalue. Meanwhile in #2 the initial argument (1) and recursive argument g(a + 1) are both int prvalue. You can check that making a change to #1 such that both recursive arguments are int prvalue (e.g. calling g(a + 1, b + 0)) will break it, while changing #2 to pass int lvalue as the recursive argument (e.g. g(++a)) will fix it.
This means that the return type deduction for the initial call is self-referential, in that it depends on the type of precisely the same call to y_combinator<lambda #2>::operator()<int>(int&&) (whereas in #1 the initial call to y_combinator<lambda #1>::operator()<int, int>(int&&, int&&) depends on y_combinator<lambda #1>::operator()<int, int&>(int&&, int&)).
Supplying the return type explicitly as in #3 means there is no self-referential type deduction, and everything is fine.
You might ask, why is #1 OK given that the recursive case is still self-referential (noting that all 3 compilers agree). This is because once we can get into the lambda's own type deduction, [dcl.spec.auto]/10 kicks in and the first return statement gives a return type to the lambda, so when it recursively calls g, that type deduction has already succeeded.
A diagram usually helps:
y_combinator<lambda #1>::operator()<int, int>
-> forwards to [lambda #1]::operator()<y_combinator<lambda #1>> {
has return type int by [dcl.spec.auto]/10
calls y_combinator<lambda #1>::operator()<int, int&> (not previously seen)
-> forwards to [lambda #1]::operator()<y_combinator<lambda #1>>
-> already deduced to return int
-> this is OK
}
y_combinator<lambda #2>::operator()<int>
-> forwards to [lambda #2]::operator()<y_combinator<lambda #2>> {
has return type int by [dcl.spec.auto]/10
calls y_combinator<lambda #2>::operator()<int>
but y_combinator<lambda #2>::operator()<int> has incomplete return type at this point
-> error
}
A fix (thanks to #aschepler) is to remember the argument lists that the lambda has been called with already, and provide a "clean" wrapper whose functional call operator(s) are not yet undergoing return type deduction for each new set of argument types:
template<class...> struct typelist {};
template<class T, class... Ts>
constexpr bool any_same = (std::is_same_v<T, Ts> || ...);
template <class F>
struct y_combinator {
template <class... TLs>
struct ref {
y_combinator& self;
template <class... Args>
decltype(auto) operator()(Args&&... args) const {
using G = std::conditional_t<
any_same<typelist<Args...>, TLs...>,
ref<TLs...>,
ref<TLs..., typelist<Args...>>>;
return self.f(G{self}, std::forward<Args>(args)...);
}
};
F f;
template <class... Args>
decltype(auto) operator()(Args&&... args) {
return ref<>{*this}(std::forward<Args>(args)...);
}
};
template <class F> y_combinator(F) -> y_combinator<F>;
Using this code:
template <class F>
struct y_combinator {
F f; // the lambda will be stored here
// a forwarding operator():
template <class... Args>
decltype(auto) operator()(Args&&... args) const {
// we pass ourselves to f, then the arguments.
// [edit: Barry] pass in std::ref(*this) instead of *this
return f(*this, std::forward<Args>(args)...);
}
};
without the std::ref (which is used for efficiency i believe, because you don't copy the object over and over) errors changes to
prog.cc:23:18: error: function 'operator()<int>' with deduced return type cannot be used before it is defined
So probably the compiler can't figure out the return type, but i can't tell you how in the first case it can deduce it

Why do I need to specify a return value for a function I'm passing to a Y combinator

I wrote a Y combinator like such:
template <class F>
struct Y{
F f;
Y(F _f) : f{_f} {}
template<class...arg_t>
auto operator()(arg_t&&...arg) {return f(*this,std::forward<arg_t>(arg)...);}
};
It works, but when I tried to define a factorial
auto fact = Y{[](auto&& self, int n) {
if (n<=1) return 1;
return n*self(n-1);}};
it would compile, but when I called it like f(3) clang got stuck on deducing the return type. With an explicit return type, it all worked fine. Is this a limitation of the template deduction? Is there a work-around?
I don't believe there is a way around it. You create a lambda with the following definition:
[](auto&& self, int n) {
if (n<=1) return 1;
return n*self(n-1);
}
This translates to:
struct lambda
{
template <typename T1>
constexpr auto operator()(T1&&self, int n) const
{
if (n<=1)
return 1;
return n*self(n-1);
}
};
Given that code, your compiler should deduce the return type as the common type of the 2 return statements.
With your template instation, it first needs to know the return type of your instantiation before it calculate the answer of that instantiation.
For this specific case, it might still be possible to deduce it correctly. What happens if you add extra indirections in between and recourse back to your type?
Type deduction is applied to the two return statements of the Y combinator, unconditionally, because the information held by the variable n is not a constant expression (an expression that is known by the compiler at compilation time). So the fixed point is not found by type deduction.
If n's value is known at compilation time, type deduction will succeed, example:
struct fact_overloads{
template<class Self,int n>
constexpr auto
operator()(Self&& self, std::integral_constant<n>){
if constexpr (n<=1) return 1;
else return n * self(std::integral_constant<n-1>{});
};
};
auto fact = Y{fact_overloads{}};
But such a function has a limited set of use cases because the value of n must be know at compil time.

Simplest way to determine return type of function

Given a very simple, but lengthy function, such as:
int foo(int a, int b, int c, int d) {
return 1;
}
// using ReturnTypeOfFoo = ???
What is the most simple and concise way to determine the function's return type (ReturnTypeOfFoo, in this example: int) at compile time without repeating the function's parameter types (by name only, since it is known that the function does not have any additional overloads)?
You can leverage std::function here which will give you an alias for the functions return type. This does require C++17 support, since it relies on class template argument deduction, but it will work with any callable type:
using ReturnTypeOfFoo = decltype(std::function{foo})::result_type;
We can make this a little more generic like
template<typename Callable>
using return_type_of_t =
typename decltype(std::function{std::declval<Callable>()})::result_type;
which then lets you use it like
int foo(int a, int b, int c, int d) {
return 1;
}
auto bar = [](){ return 1; };
struct baz_
{
double operator()(){ return 0; }
} baz;
using ReturnTypeOfFoo = return_type_of_t<decltype(foo)>;
using ReturnTypeOfBar = return_type_of_t<decltype(bar)>;
using ReturnTypeOfBaz = return_type_of_t<decltype(baz)>;
Most simple and concise is probably:
template <typename R, typename... Args>
R return_type_of(R(*)(Args...));
using ReturnTypeOfFoo = decltype(return_type_of(foo));
Note that this won't work for function objects or pointers to member functions. Just functions, that aren't overloaded or templates, or noexcept.
But this can be extended to support all of those cases, if so desired, by adding more overloads of return_type_of.
I don't know if is the simplest way (if you can use C++17 surely isn't: see NathanOliver's answer) but... what about declaring a function as follows:
template <typename R, typename ... Args>
R getRetType (R(*)(Args...));
and using decltype()?
using ReturnTypeOfFoo = decltype( getRetType(&foo) );
Observe that getRetType() is only declared and not defined because is called only a decltype(), so only the returned type is relevant.

Getting return type of an overloaded member function

I am trying to determine the return type of an overloaded member function to later use that type in my function template (see the example below). Cannot figure out how to do it using C++11 templating machinery (without modifying the definitions for struct A and B in the code below). Is this doable (in C++11 in general and MSVS2013 in particular) and how?
struct A{};
struct B{};
struct X
{
double f(A&);
int* f(B&);
};
template<typename T, typename R = /*??? what X::f(T) returns ???*/>
R ff(T& arg)
{
X x;
R r = x.f(arg); // preferably if I can create local variables of type R here
return r;
}
int main()
{
A a; ff(a);
B b; ff(b);
}
You can use decltype() for this, using std::declval to simulate values of the types needed to create the method call expression:
typename R = decltype(std::declval<X>().f(std::declval<T&>()))
Here is a demo that outputs the type ID of R; you can see that it correctly deduces double and int * for ff(a) and ff(b) respectively.
Side note: the entire body of your template function can be reduced to return X().f(arg);.
You can also use C++14 auto return type deduction like this:
template<typename T>
auto ff(T& arg)
{
X x;
auto r = x.f(arg);
return r;
}
In C++11, you can use the late return type:
template <typename T>
auto ff(T&& arg) -> decltype(std::declval<X>().f(arg))
{
return X().f(arg);
}
In C++14, you can even omit the late return type and let the compiler determine everything itself, like in Baum mit Augen's answer.
Edit: made the late return type work for non-default constructable types X.

C++11 constexpr function's argument passed in template argument

This used to work some weeks ago:
template <typename T, T t>
T tfunc()
{
return t + 10;
}
template <typename T>
constexpr T func(T t)
{
return tfunc<T, t>();
}
int main()
{
std::cout << func(10) << std::endl;
return 0;
}
But now g++ -std=c++0x says:
main.cpp: In function ‘constexpr T func(T) [with T = int]’:
main.cpp:29:25: instantiated from here
main.cpp:24:24: error: no matching function for call to ‘tfunc()’
main.cpp:24:24: note: candidate is:
main.cpp:16:14: note: template<class T, T t> T tfunc()
main.cpp:25:1: warning: control reaches end of non-void function [-Wreturn-type]
clang++ -std=c++11 says that template's parameters of tfunc<T, t>() are ignored because invalid.
Is that a bug, or a fix ?
PS:
g++ --version => g++ (GCC) 4.6.2 20120120 (prerelease)
clang++ --version => clang version 3.0 (tags/RELEASE_30/final) (3.0.1)
The parameter t is not a constant expression. Hence the error. It should be also noted that it cannot be a constant expression.
You can pass the constant expression as argument, but inside the function, the object (the parameter) which holds the value, is not a constant expression.
Since t is not a constant expression, it cannot be used as template argument:
return tfunc<T, t>(); //the second argument must be a constant expression
Maybe, you want something like this:
template <typename T, T t>
T tfunc()
{
return t + 10;
}
template <typename T, T t> //<---- t became template argument!
constexpr T func()
{
return tfunc<T, t>();
}
#define FUNC(a) func<decltype(a),a>()
int main()
{
std::cout << FUNC(10) << std::endl;
}
Now it should work : online demo
I get the feeling that constexpr must also be valid in a 'runtime' context, not just at compile-time. Marking a function as constexpr encourages the compiler to try to evaluate it at compile-time, but the function must still have a valid run-time implementation.
In practice, this means that the compiler doesn't know how to implement this function at runtime:
template <typename T>
constexpr T func(T t)
{
return tfunc<T, t>();
}
A workaround is to change the constructor such that it takes its t parameter as a normal parameter, not as a template parameter, and mark the constructor as constexpr:
template <typename T>
constexpr T tfunc(T t)
{
return t + 10;
}
template <typename T>
constexpr T func(T t)
{
return tfunc<T>(t);
}
There are three levels of 'constant-expression-ness':
template int parameter, or (non-VLA) array size // Something that must be a constant-expression
constexpr // Something that may be a constant-expression
non-constant-expression
You can't really convert items that are low in that list into something that is high in that list, but obviously the other route it possible.
For example, a call to this function
constexpr int foo(int x) { return x+1; }
isn't necessarily a constant-expression.
// g++-4.6 used in these few lines. ideone doesn't like this code. I don't know why
int array[foo(3)]; // this is OK
int c = getchar();
int array[foo(c)]; // this will not compile (without VLAs)
So the return value from a constexpr function is a constant expression only if all the parameters, and the implementation of the function, can be completed at executed at compile-time.
Recap the question: You have two functions which take a parameter of type T. One takes its parameter as a template parameter, and the other as a 'normal' parameter.
I'm going to call the two functions funcT and funcN instead of tfunc and func.
You wish to be able to call funcT from funcN. Marking the latter as a constexpr doesn't help.
Any function marked as constexpr must be compilable as if the constexpr wasn't there. constexpr functions are a little schizophrenic. They only graduate to full constant-expressions in certain circumstances.
It would not be possible to implement funcN to run at runtime in a simple way, as it would need to be able to work for all possible values of t. This would require the compiler to instantiate many instances of tfunc, one for each value of t. But you can work around this if you're willing to live with a small subset of T. There is a template-recursion limit of 1024 in g++, so you can easily handle 1024 values of T with this code:
#include<iostream>
#include<functional>
#include<array>
using namespace std;
template <typename T, T t>
constexpr T funcT() {
return t + 10;
}
template<typename T, T u>
constexpr T worker (T t) {
return t==0 ? funcT<T,u>() : worker<T, u+1>(t-1);
}
template<>
constexpr int worker<int,1000> (int ) {
return -1;
}
template <typename T>
constexpr T funcN(T t)
{
return t<1000 ? worker<T,0>(t) : -1;
}
int main()
{
std::cout << funcN(10) << std::endl;
array<int, funcN(10)> a; // to verify that funcN(10) returns a constant-expression
return 0;
}
It uses a function worker which will recursively convert the 'normal' parameter t into a template parameter u, which it then uses to instantiate and execute tfunc<T,u>.
The crucial line is return funcT<T,u>() : worker<T, u+1>(t-1);
This has limitations. If you want to use long, or other integral types, you'll have to add another specialization. Obviously, this code only works for t between 0 and 1000 - the exact upper limit is probably compiler-dependent. Another option might be to use a binary search of sorts, with a different worker function for each power of 2:
template<typename T, T u>
constexpr T worker4096 (T t) {
return t>=4096 ? worker2048<T, u+4096>(t-4096) : worker2048<T, u>(t);
}
I think this will work around the template-recursion-limit, but it will still require a very large number of instantiations and would make compilation very slow, if it works at all.
Looks like it should give an error - it has no way of knowing that you passed in a constant value as t to func.
More generally, you can't use runtime values as template arguments. Templates are inherently a compile-time construct.